GPT Image 2.5 ships as two variants, Flare and Sunburst, at the same price per image. The obvious reading is that one is the fast one and one is the good one, and you should use "good" when it matters.
These five pairs are renders I made on Callirra, where both variants are selectable side by side at the same price.
After rendering five prompts through both at identical settings, that reading is wrong. They are not a quality ladder. They are two different failure modes, and the failures are predictable from what the picture is about.
How the test was run
Five prompts, both variants, same quality (xhigh), same aspect ratio per case, no cherry-picking: the three below are what the page shows, and the other two follow the same pattern.
| Prompt | What it is about | Winner | Flare | Sunburst |
|---|---|---|---|---|
| Direct flash night portrait | Light on skin | Flare | 27 s | 49 s |
| Night market queue | Crowd, steam, colour temperature | tie | 27 s | 57 s |
| Golden hour backlight | Light itself | Flare | 24 s | 46 s |
| Retro racing replay | HUD type, five exact strings | Sunburst | 26 s | 46 s |
| Isometric dungeon key art | Structure, cutaway, layout | Sunburst (narrow) | 35 s | 57 s |
Two facts from the same table: Flare is roughly half the wall-clock, and the winner alternates — so "use the better one" is not an instruction you can follow without looking at the prompt.
What each variant does differently
Flare renders light as an event. In the night portrait it kept pores, freckles and the oily highlight across the forehead; Sunburst smoothed the face marginally. In the backlight shot Flare put a real flare streak across the frame and turned hair into a rim that reads as light rather than as a painted edge; Sunburst came back flatter and more even.
Sunburst renders arrangement as a system. On the racing replay both variants placed all five HUD strings correctly, but Sunburst stacked them into one tight block with a segmented speed bar while Flare scattered them to four corners in flatter type. On the isometric dungeon it resolved cut walls, roots and rubble more distinctly.
Translate that into a rule:
- If the picture is about light, use Flare.
- If the picture is about arrangement — type, interface, layout — use Sunburst.
Everything else is either a tie or a coin flip, and at identical prices a tie means you should pick the faster one.
The second-order cost nobody mentions
Flare finishing in 24–35 seconds versus Sunburst at 46–57 changes how you work, not just how long you wait. A 30-image batch is 15 minutes on Flare and around 27 on Sunburst. If you are exploring compositions, that is the difference between two rounds of review in an afternoon and one.
So the practical default is: explore on Flare, promote to Sunburst when the prompt is about type or layout — not when it "feels important".
How it looks in a request
curl https://api.callirra.com/v1/images \
-H "Authorization: Bearer $CALLIRRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2-5-flare",
"prompt": "one hard flash, a wall, nothing else in frame…",
"resolution": "2K",
"quality": "xhigh",
"aspect_ratio": "3:4"
}'Swap the model slug to gpt-image-2-5-sunburst and everything else stays identical. That is what makes a fair comparison possible — and what makes the twin renders above worth trusting.
Where this goes wrong
Assuming Sunburst is the quality tier. It is the arrangement tier. On a light-dominated prompt it is the worse choice and it takes twice as long.
Comparing across settings. If the resolution or quality differs between the two renders, you are measuring the settings, not the variants.
Switching mid-series. Mixing variants inside one set of product images changes the grading; it reads as two shoots.
Buying both for the same frame "to be safe". They cost the same, so compare, don't stack. Pick per prompt, not per project.
When either is fine
Flat graphic work with no type, simple product cutouts, and anything destined for a 400 px thumbnail: the difference disappears. In those cases use Flare for the speed and stop worrying about it.
FAQ
Is Sunburst higher resolution? No. Both support 1K, 2K and 4K; the resolution and quality axes are separate from the variant.
Does the variant change the price? No. Same per-image credit at the same resolution and quality.
Can I use references with either? Yes, up to 16 reference images on image-to-image.
Which one for a poster with a long headline? Sunburst, and check the letterforms at 100% before publishing.
The five measured pairs, with prompts and timings, are on Callirra's GPT Image 2.5 page. If you find a prompt where the rule inverts — arrangement-heavy and Flare wins, or light-heavy and Sunburst wins — that is the most useful thing you could send me.